hang in jampl_jac

[mlubin]

On my machine (OS X 10.7), test.jl is hanging in jampl_jac at

jl_value_t *float64_array_type = jl_apply_array_type(jl_float64_type, 1);

The backtrace is:

#0  0x000000010d067e30 in alloc_4w ()
#1  0x000000010d055422 in jl_tuple2 ()
#2  0x000000010d05baf8 in jl_apply_array_type ()

digging in more, I found that Julia gets stuck in this loop in add_page in gc.c:

while ((char*)v <= lim) {
    *pfl = v;
    pfl = &v->next;
    v = (gcval_t*)((char*)v + p->osize);
}

@JeffBezanson @Keno, any idea what could lead to this? The code in ampl.jl uses the Julia C API.

@dpo, I was going to suggest this at some point, but perhaps the C-level code could be simplified by avoiding using julia.h and instead letting the julia wrapper allocate any arrays? This would also make it easier to re-use array storage instead of allocating new vectors on every call. At the julia level, you could still have a function like hess that returns a tuple, but the tuple would be allocated in julia, avoiding any potential sources of error in interacting with the Julia C interface and GC. (It's not clear that an error in your code is causing this issue, but it would greatly simplify debugging if we could discard that possibility.)

[mlubin]

Here's the output of versioninfo():

Julia Version 0.3.0-rc1+307
Commit 60c3f40* (2014-08-01 22:08 UTC)
DEBUG build
Platform Info:
  System: Darwin (x86_64-apple-darwin11.4.2)
  CPU: Intel(R) Core(TM)2 CPU         T7400  @ 2.16GHz
  WORD_SIZE: 64
  BLAS: libopenblas (DYNAMIC_ARCH NO_AFFINITY Core2)
  LAPACK: libopenblas
  LIBM: libm
  LLVM: libLLVM-3.3

[dpo]

@mlubin I'm not seeing any hangs here, but you're right and that was on my list of things to do. Since I couldn't find any documentation, I simply followed examples to create tuples in Julia. A GC-related bug was fixed recently. It's very possible that there's still a bug in there. Now I wonder if there's a performance penalty to allocating arrays at the Julia level (though I agree that the maintenance benefit might outweigh it).

[mlubin]

Actually I can see two reasons why allocating on the Julia side would improve performance:

1. Type inference for the returned values of jac, hess, etc currently fails because Julia has no way to know that the C functions return tuples (this could be fixed by adding type assertions)
2. Arrays allocated by julia (vs. malloc) are aligned at memory boundaries for improved performance with BLAS/SIMD operations